ordlibrary/clawd-chart-foundation-training

Clawd Chart Foundation Training

A provenance-tracked text-and-vision training package for ecosystem-native Solana AI: supervised chat, chart question answering, and document adaptation. Published by ordlibrary as part of the Solana Clawd ecosystem.

Official catalog: Clawd AI · Solana Clawd · Training repository · Model Kit

Contents and verified counts

The repository contains chart-foundation-data.zip and train_chart_foundation.py. Counts below were checked against the published archive on October 1, 2026, including every JSONL record and the image file inventory.

Split	Supervised chat / chart rows	Document chunks
Train	34,831	2,338
Validation	3,257	149
Test	3,126	858
Total	41,214	3,345

The archive includes 1,712 chart images. Document chunks are not necessarily independent documents; chunks from the same document share a group and split. The supplied preparation manifest records training_completed: false: dataset publication is not evidence that a full model training run completed.

Data structure

After extraction:

train.jsonl
validation.jsonl
test.jsonl
documents-train.jsonl
documents-validation.jsonl
documents-test.jsonl
images/
manifest.json

Supervised rows have id, messages, images, sources, split, and group. messages is a list of role/content objects; images is empty for text-only examples or contains an image path relative to the extracted package. Document rows have text, source, split, and group.

Download and load

The training data is packaged inside the ZIP rather than as root-level JSONL. Download and extract it before using the JSON loader:

from pathlib import Path
from zipfile import ZipFile
from huggingface_hub import hf_hub_download
from datasets import Features, List, Value, load_dataset

archive = hf_hub_download(
    "ordlibrary/clawd-chart-foundation-training",
    "chart-foundation-data.zip",
    repo_type="dataset",
)
root = Path("clawd-chart-foundation-data").resolve()
root.mkdir(exist_ok=True)
with ZipFile(archive) as package:
    # Verify paths before extracting an externally downloaded archive.
    for member in package.infolist():
        if not (root / member.filename).resolve().is_relative_to(root):
            raise ValueError("Unsafe archive path")
    package.extractall(root)

# Early text-only batches have empty images lists. Declare their type so the
# loader does not infer a null element type before reaching vision examples.
sft_features = Features({
    "id": Value("string"),
    "messages": List({"role": Value("string"), "content": Value("string")}),
    "images": List(Value("string")),
    "sources": List(Value("string")),
    "split": Value("string"),
    "group": Value("string"),
})
sft = load_dataset("json", features=sft_features, data_files={
    split: str(root / f"{split}.jsonl")
    for split in ("train", "validation", "test")
})
documents = load_dataset("json", data_files={
    split: str(root / f"documents-{split}.jsonl")
    for split in ("train", "validation", "test")
})
# Resolve each SFT image path against root before feeding a vision processor.

Sources and preparation

The archive manifest inventories existing Clawd instruction datasets and their named held-out splits; repository and research document corpora; the ordlibrary/charts bucket's trajectories and chart metadata; captured Clawd WebSocket frames; PDF reference material; and a 33-row historical SolArchive sample. It does not contain the full SolArchive history. Source paths in the manifest record preparation provenance and need not exist on the consumer's machine.

The preparation code requires supervised conversations to contain a user and end in an assistant response. It removes exact duplicate chat records, filters recognized secret patterns, preserves explicitly named validation/test splits, and assigns remaining groups deterministically. Duplicate/group collisions are moved toward the more restrictive held-out split. Image families and chunks from the same document remain grouped. Dataset cards, manifests, FAISS indexes, and unusable preference records are not automatically treated as supervised labels.

Preparation rejects recorded in the archive manifest: 1,866 duplicate chats, 21 chat secret-pattern matches, 28 secret-bearing documents, 7 non-SFT records, and 2 invalid image metadata rows. These checks are not a comprehensive privacy audit or proof of semantic independence.

The manifest names DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU as the intended base model. Model compatibility, licensing, and vision processing must be checked against the chosen base model and training script.

Intended uses and limitations

• Research on Solana mechanics, chart understanding, and source-grounded agents.
• Text/vision supervised fine-tuning and separate document adaptation.
• Evaluation with the preserved validation and test partitions.

This is a dated training snapshot, not a realtime price feed or a trading oracle. Examples may contain synthetic charts, historical facts, or outdated provider interfaces. Exact duplicate filtering does not establish semantic separation; inspect source families before interpreting held-out results. The dataset does not establish profitability, calibration, production safety, or completed training. Preserve Brain/Hands separation: model outputs do not grant signing authority.

Licensing and attribution

The bundle combines multiple sources and does not declare a blanket license here. Review individual source rights and the intended base model's terms before reuse. The manifest attributes the SolArchive sample as CC-BY-4.0 — Data from SolArchive.org. That attribution does not extend to all other contents.

Citation

@misc{clawd_chart_foundation_training,
  author = {Solana Clawd and ordlibrary},
  title = {Clawd Chart Foundation Training},
  year = {2026},
  url = {https://huggingface.co/datasets/ordlibrary/clawd-chart-foundation-training},
  note = {Dataset card verified against the published archive on 2026-10-01}
}